Why Are Beveridge-Nelson and Unobserved-Component

Decompositions of GDP So Different?

By James C. Morley*, Charles R. Nelson**, and Eric Zivot**

*
Department of Economics, Washington University in St. Louis
**
Department of Economics, University of Washington

First draft: July 23, 1999

This draft February 23, 2001

PRELIMINARY, Not to be quoted without authors' permission.

Copyright  1999, 2000, 2001 by James C. Morley, Charles R. Nelson, and Eric Zivot

The authors acknowledge with thanks support from the National Science Foundation
under grants SES-9818789 and SBR-9711301 and from the Van Voorhis and Ensley
endowments at the University of Washington. We received helpful comments from many
individuals and particularly Charles Engel, Andrew Harvey, Chang-Jin Kim, Chris
Murray, Jeremy Piger, Simon Potter, Glenn Rudebusch, Richard Startz and Mark Wohar,
but responsibility for errors is entirely the authors'. Kyongwook Choi provided valuable
research assistance.

Address correspondence to: Charles R. Nelson, Box 353330, Univ. of Washington,

Seattle, WA 98195, USA; cnelson@u.washington.edu.

1
 Abstract

This paper reconciles two widely-used decompositions into trend and cycle that
are both model-based but yield starkly contrasting results. Beveridge-Nelson implies that
stochastic trend accounts for most of the variation in output, while Unobserved
Components implies that cyclical variation is dominant instead. Which characterization is
correct has broad implications for identifying the source of aggregate fluctuations, and
the conflicting results remain a puzzle in the literature. We show that these differ because
of the restriction usually imposed in the UC framework that trend and cycle innovations
are uncorrelated. When this restriction is relaxed the two approaches yield identical
decompositions, and the zero restriction is rejected for U.S. quarterly GDP.
JEL classification: C22, C5, E32.

2
 The decomposition of real GDP into trend and cycle remains a problem of
considerable practical importance, but two widely used methods yield starkly different
results. The unobserved component approach, introduced by Harvey (1985) and Clark
(1987), implies a very smooth trend with a cycle that is large in amplitude and highly
persistent. In contrast, the approach of Beveridge and Nelson (1981) implies that much of
the variation in the series is attributable to variation in the trend while the cycle
component is small and noisy. This conflict is apparent in Figures 1 and 2 in this paper
where the two cycle components are plotted, and has been widely noted; see Watson
(1986), Stock and Watson (1988) among others.
It should surprise us that the unobserved component (UC) and Beveridge-Nelson
(BN) methods produce very different trend-cycle decompositions since both are model-
based. Each implies an ARIMA representation. Neither imposes smoothness in the trend
component a priori as does the smoother of Hodrick and Prescott (1997) or as in the
polar case of a linear trend that forces all variation, save constant growth, into the cycle.
The UC and BN both "let the data speak for itself" in this regard. While it is often stated
that BN assumes a perfect negative correlation between trend and cycle innovations, that
is a property of the estimated trend and cycle, not the unobserved components, and it is a
property shared with the UC decomposition. This paper attempts to find out why we do
not, after decades of research, have a consistent picture of how variation in a series like
real GDP should be allocated between trend and cycle.
Briefly, section 1 demonstrates the theoretical equivalence between the
approaches. Section 2 investigates the source of the difference observed in practice.
Section 3 concludes.

3
1. Theoretical Equivalence of the Beveridge-Nelson and Unobserved Component
Estimates of Trend and Cycle
The detrending problem is motivated by the idea that the log of aggregate output
is usefully thought of as the sum of a nonstationary trend component that accounts for
long-term growth and a stationary component that allows for transitory deviations of
output from trend. We follow custom in referring to the latter as the “cycle” even if it is
not periodic. The UC model takes the form:

(1a) yt = τt + ct
(1b) τt = τt-1 + µ + ηt; η ~ i.i.d. N(0, σ η2 )

(1c) ct is stationary and ergodic

where {yt} is the observed series, {τt} is the unobserved trend assumed to be a random
walk with average growth rate µ, and {ct} is the unobserved stationary cycle component.1
The UC-ARMA adds the condition that {ct} is a stationary and invertible ARMA(p,q)
process with innovations that may be contemporaneously cross-correlated with trend
innovations,

(1d) φp(L) ct = θq(L) εt; ε ~ i.i.d. N(0, σ ε2 ); Cov(εt, ηt+k) = σεη for k=0; 0 otherwise.

In some implementations the rate of drift µ is allowed to evolve as a random walk,

although this has little influence on the estimated cycle component for U.S. GDP, and an
additional irregular term is sometimes added. Harvey (1985) and Clark (1987), and
Harvey and Jaeger (1993) suggest specifying p=2 which allows the cycle process to be
periodic in the sense of having a peak in its spectral density function. They and others
further assume that the trend and cycle innovations are uncorrelated, setting

1
As noted by Blanchard and Quah (1989), the structural trend of output does not necessarily follow a
random walk. Therefore, a cycle component that represents the deviation from a random walk trend may
reflect both transitory effects of supply shocks and cyclical movements due to demand shocks.

4
(1e) σεη = 0,

thereby casting the UC model in state-space form by treating (1a) as the measurement
equation and (1b) as the state transition equation, see the technical appendix for details.
We denote this constrained zero-covariance UC-ARMA model as UC-0.
In practice, the parameters are unknown and are estimated from the data series
(y1, .., yn) by putting the model in state-space form and using the maximum likelihood
method of Harvey (1981) based on the prediction error decomposition generated by the
Kalman Filter. Given the parameters, the Kalman filter is used to compute the
expectation of the stochastic trend component conditional on data through time t:

τ t t = E [τ t Yt ] , where Yt = ( y1 ,.., y t ) .
)

Alternatively, the BN estimate of stochastic trend for an I(1) time series {yt} is
defined to be the limiting forecast as horizon goes to infinity, adjusted for the mean rate
of growth; so

BN t = lim E [y t + M − Mµ Yt ].
M →∞

BN showed that the time series {BNt} will be a random walk with drift, the deviation
from trend is a stationary process, and that the innovations of {BNt} and {yt - BNt} are
perfectly correlated. The series {BNt} is typically calculated from an estimated ARIMA
representation of {yt}, which in principle is unique after cancellation of any redundant
AR and MA factors2.
It is well known that the UC–ARMA model always implies a univariate ARIMA
representation for {yt}. This is what Nerlove, Grether, and Carvalho (1979) refer to as the

2
The theoretical justification for the BN decomposition and its relationship to Martingale decompositions
is given in Phillips and Solo (1992). A corresponding decomposition for seasonal time series is given in
Box, Pierce, and Newbold (1987).

5
canonical form of the UC model, and it may be useful to think of it as the reduced form.
Substituting (1b) and (1d) into (1a), taking first differences, and rearranging we obtain

(2a) φp(L) (1-L)yt = φp(1)µ + φp(L) ηt + θq(L) (1-L)εt .

Recognizing that the right hand side will have non-zero autocorrelations through lag
max(p, q+1), Granger’s Lemma (see Granger and Newbold (1986) pg. 29) implies that
the univariate representation will be

(2b) φp(L) (1-L)yt = µ∗ + θ q** ( L) ut; u ~ i.i.d. N(0, σ u2 ); q* = max(p, q+1)

where the coefficients of θ q** ( L) and σ u2 are obtained by matching the autocovariances

of the right hand side of (2a) and (2b); see Watson (1986). This ARIMA reduced form
fully describes the joint distribution of the {yt} and therefore the conditional distribution
of future observations given the past and is unique. Note that the BN trend for the
reduced-form ARIMA model (2b) may be derived from the Wold representation of (2b)
and expressed as
t
(2c) BN t = BN t −1 + ψ (1)ut = ψ (1)∑ u j ,
j =1

where ψ(1) = θ q** (1) /φp(1) and BN0 = 0. From (2a) it can be seen that the variance of the

innovation to the BN trend is ψ (1) 2 σ u2 . The BN cycle is obtained by subtracting from yt

the BN trend and other deterministic components.

Correspondingly, there is always at least one UC representation of any given
ARIMA process; as Cochrane (1988) pointed out, the existence of the BN decomposition
guarantees this. In general, however, there will not be a unique UC representation since
the parameters may not all be identified. For example, consider the ARIMA(0,1,1)
process so that in the notation of (2b) the orders are p=0 and q*=1. By inspection of (2a)
it is clear that this implies q=0, hence the ARIMA(0,1,1) process has a UC representation

6
as a random walk plus noise, and the first differences are autocorrelated at lag one only.
However, the parameters of the UC representation of the ARIMA(0,1,1) process are not
all separately identified. This is easy to see from the relation between the two non-zero
autocovariances γ j at lags 0 and 1 (values of which can be inferred from data), and the

UC parameters as follows:

γ 0 = σ η2 + 2σ ε2 + 2σ εη
γ 1 = −σ ε2 − σ εη
γ j = 0, j ≥ 2

While there are three parameters in this case, there are only two pieces of information.
Note, however that the variance of the trend innovations may be inferred by adding 2γ1 to
γ0, while the variance of the cycle innovations and the covariance between them are not
separately identified, only their sum is. This reflects a basic theme of this paper: the trend
is identified from the univariate properties of the series though the cycle may not be. It is
also possible to infer some inequality restrictions; see the discussion in Nelson and
Plosser (1981). It is easily shown that there will be at least as many non-zero moment
relations as parameters if p ≥ q + 2 , a fact which we use later to identify the covariance
in the Harvey-Clark UC model. However, the fact that the covariance is not identified in
a case as familiar as the random walk plus noise may well have contributed to the
impression evident in the literature that zero covariance is always an identifying
restriction for the UC model.
Given that a time series will not in general have a unique UC representation, it
seems surprising to us that the BN trend is the conditional expectation of the random
walk component of an I(1) process. As pointed out by Watson (1986), this is true
regardless of the covariance structure of the unobserved components. To see this,
consider the unconstrained UC model defined by (1a)-(1c), so cycle and trend
innovations may be cross-correlated. The conditional expectation of the trend component
at time t is

7
 E [τ t Yt ] = E [τ t + ct + M Yt ]

for large enough M, since the cycle, by its ergodicity, has expectation zero far enough in
the future. Further, the expected value of any future innovation in the trend is zero, so we
have

 
E [τ t Yt ] = E τ t + ∑ηt + j + ct + M Yt  .
M

 j =1 

Recognizing that the terms of the right include all the elements of yt+M except the
accumulated drift, we have

 
E [τ t Yt ] = E τ t + ∑ηt + j + ct + M Yt  = E [y t + M − Mµ Yt ] = BN t .
M

 j =1 

Then the conditional expectation of the cycle at time t is simply

E [ct Yt ] = y t − E [τ t Yt ] = yt − BN t .

Thus, we can always compute conditional expectation estimates of trend and

cycle at any point in time from the ARIMA representation of the observed series. The
two assumptions, (a) the trend is a random walk, and (b) the cycle is ergodic, are
sufficient to identify the components, and this does not depend on knowing the
covariance between trend and cycle innovations. Intuitively, the forecast at a long enough
horizon reflects only the permanence of the random walk trend. Stronger assumptions
may be needed to identify the parameters of a UC representation, but they are irrelevant if
the only objective is to estimate trend and cycle. In that case, only the conditional

8
expectation of the future given the past is required, and the ARIMA reduced form
provides the relevant conditional distribution.
It follows that if a particular time series does have a representation as a UC-0
process, then the Kalman filter and BN estimates of trend and cycle will be the same, as
long as the parameters of the ARIMA model are those implied by the UC-0
representation. In that case, BN is just another way to compute τˆt t and ĉt t ; the time

series of those estimates will be the same and they will have the same properties. In
particular, their innovations will both be functions of the innovation in yt since it is only
new information that will cause the long range forecast, the trend, to change. Thus, UC-0
and BN share the often-noted property of the BN decomposition, that the innovations of
the estimated trend and cycle series are perfectly correlated, as discussed in the paper.
Further, none of these results depend on limiting UC representations that might be
modeled to the constrained UC-0 case; the corresponding ARIMA representation will
always be an equivalent way of obtaining the information relevant to estimating the trend.
To sum up this section, we have shown that whether one uses the UC approach to
trend-cycle decomposition based on a state-space representation and the Kalman filter, or
the BN approach based on long-range forecasts from a univariate ARIMA model, the
specific results should be the same. UC and BN are simply alternative ways of
calculating the same conditional expectation of the unobserved trend and cycle at a point
in time. The fact that the two have produced such different estimates of trend and cycle in
practice implies, then, that they must be based on conflicting representations of the data.
Identifying the source of the conflict is the subject of the next section.

2. In What Way Do UC and ARIMA Models of U.S. Real GDP Conflict?

The results of Section 1 imply that the differing results obtained in practice must
be traceable to restrictions on the reduced form ARIMA model implied by the UC
approach that are in conflict with the unrestricted ARIMA model used in the BN
approach. Those restrictions presumably arise from the restrictions that have been placed
on the UC–ARMA representation used in implementing the Kalman filter, the particular

9
ARMA(p, q) form of the cycle process and the zero correlation between trend and cycle
innovations, what we called the UC-0 model. Following Clark (1987) we set p=2, to
allow for cyclical dynamics, and q=0 in the UC-0 model and obtain for real GDP 1947:1
– 1998:2, in logs and multiplied by 100, the filtered estimate of ct |t shown in Figure 1.3

Confirming results in the literature, the estimated cycle produced by the UC-0
model is both large in amplitude and very persistent.4 It agrees reasonably well with the
NBER dating of the business cycle, although it leads the NBER dating at peaks.
Reflecting the smoothness of the trend, the cycle shown here is qualitatively similar to
that obtained by simple linear detrending of log output by least squares. For example,
both imply that the economy has been below trend throughout the 1990s.
Table 1 reports the maximum likelihood estimates of the parameters and their standard
errors for the UC-0 model. Detail of the estimation technique used is given in the
technical appendix. The roots of the estimated autoregressive polynomial are complex,
implying that the business cycle has a period of almost 8 years with a standard deviation
of about 3 percentage points around trend, confirming the visual impression of
persistence, periodicity, and amplitude in Figure 1. In contrast, the trend process
innovation has a standard deviation of only about 0.7 percentage points.

3
The data series used is gdpq from the DRI databank. Clark (1987) allowed the drift parameter to evolve as
a random walk, but estimates of the variance are small. We have assumed that this parameter remains
constant, implying that output is I(1).
4
The smoothed estimate of the cycle is usually presented, but presents qualitatively the same picture.

10
 6

-2

-4

-6
50 55 60 65 70 75 80 85 90 95

Fig. 1: UC-0 Cycle, U.S. Real GDP.

Percent deviation from trend, NBER recessions shaded

Table 1: Maximum Likelihood Estimates of UC-0 Parameters

Estimate Standard Error

Trend process
Drift: µ 0.8119 (0.0500)
Innovation: ση 0.6893 (0.1038)

Cycle process
φ1 1.5303 (0.1012)
φ2 -0.6097 (0.1140)
Innovation: σε 0.6199 (0.1319)

AR Roots (inverted) 0.7652 ± 0.1558i

Implied cycle: period 7.7 years, standard deviation .03.

Log Likelihood -286.6053

11
 The reduced form ARIMA representation for the UC-0 model corresponding to
(2b) is obtained as follows. Taking first differences gives

∆y t = (1 − L)τ t + (1 − L)ct
.
= µ + ηt + (1 − L)(1 − φ1 L − φ 2 L2 ) −1 ε t

Next, multiply both sides by (1 − φ1 L − φ2 L2 ) to give

(3) (1 − φ1 L − φ 2 L2 )∆y t = µ * + η t − φ1ηt −1 − φ 2ηt −2 + ε t − ε t −1 = µ * + ut + θ 1*ut −1 + θ 2*ut − 2 .

The result in (3) uses the fact that the right-hand side has a representation as an MA(2) by
Granger’s Lemma with the univariate innovations ut being i.i.d. N(0, σ u2 ), and µ* is µ(1-

φ1-φ2). It is important to recognize that the assumption σεη=0 places complicated

nonlinear restrictions on the parameters of the ARIMA(2,1,2) model (3). In particular,
Lippi and Reichlin (1992) show that the long-run persistence measure, ψ(1) = θ(1)*/φ(1),
will be less than or equal to one. Proietti and Harvey (2000) give further restrictions on
the autoregressive parameters. These restrictions are testable implications of the UC-0
model but in empirical work they are almost never tested.
While the reduced form of the UC-0 model is a restricted ARIMA(2,1,2), when
we estimate the unrestricted form of that model by exact maximum likelihood and
compute the BN cycle component from it we get the very different results seen in Figure
2. As reported in the literature, the estimated BN cycle is small in amplitude compared to
the UC-0 cycle and much less persistent. Table 2 reports the maximum likelihood
estimates of the parameters for the unrestricted reduced-form ARIMA(2,1,2) model.
Confirming the visual impression from Figure 2, the period of the cycle implied by the
AR parameters here is much shorter, only 2.4 years instead of 8. The fact that the value of
the log likelihood is greater by roughly 2 for the unrestricted ARIMA must reflect
restrictions in the UC-0 model not imposed in the reduced form, in particular the zero

12
correlation between trend and cycle innovations. Another indication that the zero
correlation restriction may not be supported by the data is that the estimated value of
persistence, ψ(1), is greater than one. To see what correlation is implied by the ARIMA
parameters, we next solve for the parameters of the unrestricted UC model of equations
(1a)-(1d) that correspond to the estimated unrestricted ARIMA parameters.

13
 2

-1

-2
50 55 60 65 70 75 80 85 90 95

Fig. 2: Beveridge-Nelson Cycle, U.S. Real GDP.

Percent deviation from trend, NBER recessions shaded.

Table 2: Maximum Likelihood Estimates for ARIMA(2,1,2)

Estimate Standard Error

Drift µ 0.8156 (0.0864)
φ1 1.3418 (0.1519)
φ2 -0.7059 (0.1730)
θ1 -1.0543 (0.1959)
θ2 0.5188 (0.2250)
SE of Regression 0.9694 (0.0478)
ψ(1) 1.2759 (0.1543)

AR roots (inverted) 0.6709 ± 0.5057i

Implied cycle: period 2.4 years
MA roots (inverted) 0.5271 ± 0.4908i

Log Likelihood -284.6507

14
 First note that the AR parameters are the same in both the UC and ARIMA
reduced form since the AR polynomial on the left side of (2) is the AR polynomial of the
UC cycle. Now the observable moments on the MA side of (2) are the mean, which
identifies µ, and the autocovariances:

γ 0 = (1 + φ12 + φ22 )σ η2 + 2σ ε2 + 2(1 + φ1 )σ ηε

γ 1 = −φ1 (1 − φ2 )σ η2 − σ ε2 − (1 − φ2 + φ1 )σ ηε
(4)
γ 2 = −φ2 σ η2 − φ2σ ηε
γ j = 0, j ≥ 3

The autocovariances on the left-hand-side of (4) are

γ 0 = σ u2 (1 + θ12 + θ 22 )
γ 1 = σ u2θ1 (1 + θ 2 )
γ 2 = σ u2θ 2 .
The system of equations in (4) can be written in matrix form as

 γ 0   1 + φ1 + φ 2 2 2(1 + φ1 )  σ η2 
    2 
 γ 1  =  − φ1 (1 − φ 2 ) − 1 − (1 − φ 2 + φ1 )  σ ε 
γ   − φ2 − φ2  σ 
 2  0  ηε 
or
γ = Φσ .
Assuming Φ is invertible, of which a necessary condition is φ2≠0, we can solve for the
UC-0 parameters as
(5) σ = Φ −1γ .
Hence, the three non-zero autocovariance for the MA(2) are just sufficient to identify the
three remaining parameters of the UC representation, namely σ η2 , σ ε2 , and σ ηε . We note

that in a particular case the solution to (4) might not imply a positive definite covariance
matrix for the trend and cycle innovations, in which case there would not exist a
corresponding UC-ARMA(2,0) representation.

15
 Table 3 compares the estimates from Table 1 for the UC-0 model with the implied
estimates from the unrestricted ARIMA(2,1,2) reduced form obtained from (5). While the
parameters for the cycle component are somewhat similar, the unrestricted reduced form
implies a standard deviation for the trend innovation that is almost twice as large
compared to the UC-0 estimate, and a correlation between trend and cycle innovations
that is large and negative instead of zero. It is important to note that the correlation
presented in Table 3 is the estimated correlation between unobserved innovations, not the
correlation between innovations in the observed estimated series τˆt t and ĉt t . The UC-0

model restricts the correlation between unobserved innovations to be zero, while the
unrestricted ARIMA reduced form estimates that correlation implicitly. Thus, the
difference in correlation estimates seen in Table 3 reflects a difference between the two
models, not a difference in detrending methods.
In contrast, innovations in the estimated components τˆt t and ĉt t are perfectly

negatively correlated regardless of whether they are computed from a UC representation

using the Kalman filter or from the corresponding reduced form ARIMA representation
using BN. Recall from section 1, that the estimated components obtained by BN and
Kalman filter procedures are numerically equivalent if the underlying models have the
same reduced form, implying that the well-known perfect negative correlation of
estimated BN components is a property shared by UC filtered estimates. Another way to
see why UC estimates have the perfect negative correlation property is to inspect the
updating equations for the Kalman filter (see Harvey (1981, chapter 4)), noting that the
random walk property of the trend component implies that the innovation in the estimated
trend component is proportional to the forecast error in predicted yt, just as for the BN
estimate of trend. As Wallis (1995) has pointed out, the distinction between the
assumptions of a model (in the UC-0 case that the components are uncorrelated) and the
properties of estimates (estimated components are correlated) is perfectly consistent with
least squares estimation, but nevertheless has been the source of frequent confusion in the
literature.

16
 Table 3: Parameters of UC-0 Model and Those Implied by Unrestricted
ARIMA(2,1,2) Reduced Form
UC Model
UC-0 Model Implied by ARIMA
Trend process
Drift: µ 0.8119 0.8156
Innovation: ση 0.6893 1.2368

Cycle process
φ1 1.5303 1.3418
φ2 -0.6098 -0.7059
Innovation: σε 0.6199 0.7487

Covariance σ εη zero (constrained) -0.8391

Correlation ρ εη zero (constrained) -0.9062

The fact that σ εη is identified in this case implies that we can relax the constraint

that it is zero in the UC model and estimate it directly by maximum likelihood. The
resulting unconstrained UC-ARMA(2,0) model, denoted UC-UR, is cast in state-space
form by including the cycle component along with the trend in the state equations, as
noted by Canova (1998) and Nelson and Kim (1999b); see appendix for details. More
generally, the order condition for identification of the unconstrained UC-ARMA(p,q)
model (including the cross-correlation between trend and cycle innovations), in the sense
of having at least as many moment equations as parameters, is p≥q+2 and it is just
satisfied in this case with p=2, q=0. Intuitively, when this holds, increasing p increases
the number of moment equations corresponding to (4) on the MA side of the univariate
ARIMA representation without increasing the number of parameters to solve for, since
those are identified by the AR side.
Figure 3 displays the filtered estimate ct |t of the transitory component for the

unrestricted UC-UR model. The estimated cycle is essentially identical to the estimated

17
cycle from the BN decomposition.5 This verifies that the filtered estimates from the UC
model and the BN estimates are equivalent.

-1

-2
50 55 60 65 70 75 80 85 90 95

Figure 3 – UC-UR Cycle (NBER dated recessions shaded)

Table 4 reports the maximum likelihood estimates of the parameters for the
unconstrained UC model. The unconstrained model was formulated to allow estimation
of either the covariance, σεη, or the correlation, ρεη, and both estimations produce
numerically the same results and both are strongly negative. A striking feature of these
estimates is that they are all essentially the same as the implied estimates from the
unrestricted ARIMA model reported in Table 3. Note also that the estimated variance of
the permanent component, σˆη2 = 1.5296 , is essentially equal to the variance of the

innovation to the BN trend from the unrestricted ARIMA(2,1,2) model,

σˆ u2ψˆ (1) 2 = 1.5297 . While the estimated value of φ2 is several times its standard error,

5
The only difference is for the first observation, which is different due to the need to provide an initial
guess for the value of the random walk trend in estimation via the Kalman filter.

18
supporting the p=2 specification, we note that this is not a standard testing situation since
the model is not identified if the null hypothesis that φ2=0 is true.
The log likelihood value for the UC-UR model is also the same as for the ARIMA
model, and significantly larger than for the restricted UC-0 model, thus confirming
identification of the covariance between trend and cycle innovations. The likelihood ratio
statistic for testing the restriction ρεη = 0, which may be interpreted as an
overidentification test, is 3.909, with a corresponding p-value of 0.048. As a check on the
small sample properties of this test, particularly whether this test rejects zero correlation
too often, we generated data from the UC model calibrated to the UC-0 estimates, tested
the null hypothesis ρεη = 0, and found the size to be approximately correct. Thus, we can
strongly reject the restriction of a zero correlation between permanent and transitory
shocks by comparing the results for the UC-0 model with either the results for the
reduced-form ARIMA model or the unrestricted UC model.
Note that the estimate of ρεη is –0.906 with an estimated standard error of 0.073,
so a Wald-type t-test implies a far smaller p-value than the likelihood ratio test reported
above. Since the estimated value of ρεη is near the boundary of admissible values the
small estimated standard error might give a misleading impression of precision. To
reconcile the two tests, we estimated the UC-ARIMA model for fixed values of ρεη = -
0.95, -0.9, .., 0.9, 0.95 and, for each model computed the likelihood ratio statistic for the
hypothesis that ρεη is equal to the imposed value. Figure 4 gives a plot of these likelihood
ratio statistics as a function of the hypothetical ρεη. The horizontal line indicates the 95
percent quantile from the chi-square distribution with 1 degree of freedom. The shape of
the plot clearly indicates a global maximum of the likelihood at the estimated value of ρεη
and the sharpness of the likelihood around that point is reflected in the small standard
error. The implied 95 percent confidence interval for ρεη obtained by “inverting” the
likelihood ratio statistic is fairly wide but excludes ρεη = 0. Thus, the difference between
the Wald and likelihood ratio test results is traced to local verses global behavior of the
likelihood function.

19
 Table 4: Maximum Likelihood Estimates for the UC-UR Model

Estimate Standard Error

Trend process
Drift: µ 0.8156 (0.0865)
Innovation: ση 1.2368 (0.1518)

Cycle process
φ1 1.3419 (0.1456)
φ2 -0.7060 (0.0822)
Innovation: σε 0.7485 (0.1614)

Roots of AR process 0.6710 + 0.5058i

0.6710 - 0.5058i

Covariance: σεη -0.8389 (0.1096)

Correlation: ρεη -0.9063 (0.0728)

Log Likelihood Value -284.6507

4
LR statistic

3 5% chi-square
critical value

0
-1 -0. -0. -0. -0. -0. -0. -0. -0. -0. 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
9 8 7 6 5 4 3 2 1
ρεη , 0

Figure 4 – Likelihood Ratio Statistics for testing ρεη=ρεη,0

20
3. Summary and Conclusions
We have shown that trend-cycle decompositions based on unobserved component
models cast in state-space form and on the long run forecast implied by an ARIMA
model differ not because they differ in principle but because the underlying empirical
models differ. In particular, the restriction that innovations in the unobserved trend and
cycle are uncorrelated has been imposed in the former but in the latter. We note that
when this restriction is relaxed in the state-space model, the two approaches lead to
identical trend-cycle decompositions and identical univariate representations. Further, the
restriction of zero correlation is strongly rejected by the data for U.S. real GDP, quarterly
1947-1998.
If we accept the implication that innovations to trend are strongly negatively
correlated with innovations to the cycle, then the case for the importance of real shocks in
the macro economy is strengthened. As pointed out by Stock and Watson (1988) in their
influential paper, real shocks tend to shift the long run path of output, so short term
fluctuations will largely reflect adjustments toward a shifting trend if real shocks play a
dominant role. For example, a positive productivity shock, such as the invention of the
Internet, will immediately shift the long run path of output upward, leaving actual output
below trend until it catches up. This implies a negative contemporaneous correlation
since this positive trend shock is associated with a negative shock to the cycle. In
contrast, a positive nominal shock, say a shift in Fed policy towards stimulus will be an
innovation to the cycle without any impact on trend.
Closing with a few caveats, we note that the decompositions considered here
share a common restriction, that the cycle process is symmetric. Recent business cycle
research suggests that asymmetry has been an important feature of postwar U.S.
experience - recessions being characterized as an occasional sharp drop followed by more
gradual recovery; see Neftci (1984), Sichel (1993, 1994), Beaudry and Koop (1993), and
Kim and Nelson (1999a). The inference that variation in GDP is dominated by variation
in trend may reflect primarily the long periods of expansion when actual output is
relatively close to potential and any cycle is short-lived and small in amplitude. Finally,

21
the decompositions considered here are univariate with only two sources of shocks, and
additional information introduced in a multivariate setting may affect estimates of trend
and cycle.

22
 References

Ariño, M. A. and P. Newbold, 1998, “Computation of the Beveridge-Nelson

Decomposition for Multivariate Economic Time Series,” Economics Letters, 61, 37-
42.

Beaudry, P. and G. Koop, 1993, “Do Recessions Permanently Change Output?” Journal
of Monetary Economics, 31, 149-163.

Beveridge, S. and C. R. Nelson, 1981, “A New Approach to Decomposition of Economic

Time Series into Permanent and Transitory Components with Particular Attention to
Measurement of the Business Cycle,” J. of Monetary Economics, 7, 151-174.

Blanchard, O., and D. Quah, 1989, “The Dynamic Effects of Aggregate Demand and
Suppply Disturbances,” American Economics Review, 79, 655-73.

Box, G. E. P., Pierce, David A., and P. Newbold, 1987, “Estimated Trend and Growth
Rates in Seasonal Time Series, ” Journal of the American Statistical Association, 82,
276-282.

Canova, F., 1998, “Detrending and Business Cycle Facts,” Journal of Monetary
Economics, 41, 475-512.

Clark, P. K., 1987, “The cyclical component of U.S. economic activity,” The Quarterly
Journal of Economics, 102, 797-814.

Cochrane, J. H., 1988, “How Big Is the Random Walk in GNP?,” Journal. of Political
Economy, 893-920.

Granger, C.W.J., and P. Newbold, 1986, Forecasting Economic Time Series, Second
Edition, Orlando: Academic Press.

Hamilton, J.D., 1994, Time Series Analysis, Princeton: Princeton University Press.

Hamilton, J.D., 1993, “State-Space Models,” in R. Engle and D. McFadden, eds.,

Handbook of Econometrics, Vol. 4, New York: North Holland.

Harvey, A. C., 1981, Time Series Models, Oxford: Philip Allen.

Harvey, A. C., 1985, “Trends and Cycles in Macroeconomic Time Series,” Journal of
Business and Economic Statistics, 3, 216-27.

23
Harvey, A. C., and A. Jaeger, 1993, “Detrending, Stylized Facts and the Business Cycle,”
Journal of Applied Econometrics, vol. 8 no. 3, July-September, 231-47.

Hodrick, R. J., and E. C. Prescott, 1997 (1981), “Postwar U.S. Business Cycles: An
Empirical Investigation,” Journal of Money, Credit, and Banking, vol. 29 no. 1,
February, 1-16.

Kim, C-J. and C. R. Nelson, 1999a, “Friedman’s Plucking Model of Business

Fluctuations: Tests and Estimates of Permanent and Transitory Components,”
Journal. of Money Credit and Banking, 31, 317-34.

Kim, C.-J., and C.R. Nelson, 1999b, State-Space Models with Markov Switching:
Classical and Gibbs Sampling Approaches, Cambridge: MIT Press.

Lippi, M., and L. Reichlin, 1992, “On Persistence of Shocks to Economic Variables. A
Common Misconception,” Journal of Monetary Economics, 29, 87-93.

Morley, J.C., 2001, “A State-Space Approach to Calculating the Beveridge-Nelson

Decomposition,” Manuscript, Washington University.

Neftci, S. N., 1984, “Are Economic Time Series Asymmetric Over the Business Cycle?”
Journal of Political Economy, vol. 92, no. 2, 307-328.

Newbold, P., 1990, “Precise and Efficient Computation of the Beveridge-Nelson

Decomposition of Economic Time Series,” Journal of Monetary Economics, 26, 453-
457.

Nerlove, M., D.D. Grether, and J.L. Carvalho, 1979, Analysis of Economic Time Series –
A Synthesis, New York: Academic Press.

Proietti, T., and A. Harvey, 2000, “A Beveridge-Nelson Smoother,” Economics Letters,

67, 139-146.

Phillips, P.C.B., and V. Solo, 1992, “Asymptotics for Linear Processes,” Annals of
Statistics, 20, 971-1001.

Sichel, D. E., 1993, “Business Cycle Asymmetry: A Deeper Look,” Economic Inquiry,
April, 224-236.

Sichel, D. E., 1994, “Inventories and the Three Phases of the Business Cycle,” Journal of
Business and Economic Statistics, vol. 12, no. 3, 269-77.

Stock, J H., and M. W. Watson, 1988, “Variable Trends in Economic Time Series,”
Journal. of Economic Perspectives, 2, 147-74.

24
Wallis, K. F., 1995, “Discussion of 'Trend Extraction: A Practitioners Guide' by M. R.
Wickens,” HM Treasury Working Paper No. 67.

Watson, M W., 1986, “Univariate detrending methods with stochastic trends,” Journal. of
Monetary Economics, 18, 49-75.

25
 Appendix
This appendix contains notes on how to calculate the BN decomposition for an
ARIMA model and how to set up the state-space model in order to estimate the UC1
model via the Kalman filter.

BN decomposition
Numerous practical methods for calculation of the exact BN decomposition have
been proposed in the literature, including Newbold (1990) and Ariño and Newbold
(1998). As discussed in Morley (2001), a particularly convenient way to calculate the
exact BN decomposition for an ARIMA model is to first convert the model into its state-
space companion form. For the ARIMA(2,1,2) model, this is given as follows:

 ∆y t − µ  φ1 φ2 θ1 θ2   ∆y t −1 − µ  1
∆y − µ  1 0 0 0  ∆y t − 2 − µ  0
 t −1 =    +  e , (A.1)
 et  0 0 0 0   et −1  1 t
      
 e t −1  0 0 1 0   et − 2  0

or, more compactly,

βt = Fβt −1 + v t (A.1')

when the eigenvalues of F have modulus less than 1 and there are no canceling roots so
that F is invertible. Then, the BN trend can be calculated at any given point of time as

BN t = y t + [1 0 0 0]F ( I − F ) −1 β t|t , (A.2)

where βt |t = E t [ βt ] can be obtained from the Kalman filter used in exact maximum

likelihood estimation of the ARIMA(2,1,2) model. We define the BN cycle as the gap
between real GDP and the BN trend, ( y t − BN t ).

26
The State-Space Model
One possible explanation for prevalence of the unnecessary zero correlation
assumption in empirical work is the way the state-space model for a UC model is
traditionally set up and estimated. In particular, the random walk trend is often treated as
the state variable, while the AR(2) cycle is treated as a residual in the observation
equation; for example, see Clark (1987). Given this setup, estimation via the Kalman
filter requires an assumption of independence between trend and cycle innovations.
However, as demonstrated in Kim and Nelson (1999), there are other ways of
setting up the state-space representation for a UC model. One possibility is to make the
observation equation an identity, with both the trend and cycle treated as state variables.
This is the approach we take:

Observation Equation:
 τt 
 
[ ]
y t ≡ 1 1 0  ct  , (A.3)
ct −1 

or, more compactly,

y t = Hβt , (A.3’)

State Equation:
 τ t   µ  1 0 0  τ t −1  1 0
         ηt 
c =
 t     0 + 0 φ1 φ c
2   t −1  + 0 1  ε  , (A.4)
ct −1   0  0 1 0  ct − 2  0 0  t 

or, more compactly,

27
 βt = µ~ + Fβt −1 + v t . (A.4’)

Then, we can use the Kalman filter to estimate the model, even if we allow the trend and
cycle innovations to be correlated. That is,

 σ η2 σ ηε 0
 
Q ≡ E[v t v t′ ] = σ ηε σ ε2 0 . (A.5)
 0 0 0

The Kalman filter is given by the following six equations:

βt |t −1 = µ~ + Fβt −1|t −1 , (A.6)

Pt |t −1 = FPt −1|t −1 F ′ + Q , (A.7)

y t − y t |t −1 = y t − Hβt |t −1 , (A.8)

f t |t −1 = HPt |t −1 H ′ , (A.9)

βt |t = βt |t −1 + K t ( y t − y t |t −1 ) , (A.10)

Pt |t = Pt |t −1 − K t HPt |t −1 , (A.11)

where βt |t −1 ≡ E[ βt |ψ t −1 ] , for example, is the expectation of βt conditional on

information up to time t − 1 ; Pt |t −1 is the variance of βt |t −1 ; f t |t −1 is the variance of

6
( yt − yt |t −1 ) ; and K t ≡ Pt |t −1 H ′f t |−t 1−1 is the Kalman gain.

6
For a more general discussion of the Kalman filter and state-space models, as well as details on the
derivation of the Kalman gain, refer to Hamilton (1993, 1994) and Kim and Nelson (1999).

28
 Given some initial values β0|0 and P0|0 (as discussed below), we can iterate through

(A.6)-(A.11) for t = 1,..., T to obtain filtered inferences about βt conditional on

information up to time t. Also, as a by-product of this procedure, we obtain ( yt − yt |t −1 )

and f t |t −1 , which we can use to find maximum likelihood estimates of the hyper-

parameters based on the prediction error decomposition:

1 T 1 T
max l (θ ) = −
θ
∑
2 t =τ +1
ln( 2πf t |t −1 ) − ∑
2 t =τ +1
( y t − y t|t −1 ) 2 f t|−t 1−1 , (A.12)

where θ = ( µ , φ1 , φ 2 , σ η , σ ε , σ ηε )' and τ gives the number of start-up values skipped in

the evaluation of the log-likelihood. We set τ = 2.

In terms of the initial values β0|0 and P0|0 , we specify the initial value of the data

for the random walk component, but assign it an extremely large variance. For the
transitory component, we use the unconditional mean and variance of the AR(2) process.
In the process of the maximizing the log-likelihood function we impose stationarity
constraints on the autoregressive parameters and a positive definiteness constraint on the
innovation covariance matrix.

29